Use of Artificial Intelligence & Machine Learning with Deep Learning for Glaucoma Detection in Human Eyes & its Real Time Hardware Implementation

—In this paper, a brief introduction to AI, ML and the Eye w.r.t. Deep Learning for Glaucoma Detection and Hardware Implementation is being presented. The result is the outcome of the Post-Graduate project work of the student that is going to be carried out in the second year of the course & this work is just the synopsis that is being framed for the carrying out of the detection of glaucoma disease.


I. INTRODUCTION
Glaucoma harms the optic nerve which prompts changeless visual deficiency. It can't be restored, so distinguishing the illness in time is significant. Glaucoma is one of the most extreme eye infections as indicated by the quantity of visual impairment causes in India and western nations and is the second most driving eye sickness. Along these lines, the early discovery, long haul observing of the patients and the choice about the suitable treatment at the right time are the genuine assignments for the ophthalmologist. This prior location of fatal illnesses has been proposed utilizing propelled picture handling, investigation and acknowledgment strategies. This condition of workmanship systems had just been helped specialists in different fields, for example, prior recognition and determination of sicknesses, clinical choices, remote detecting medical procedures, etc. In short to state, glaucoma is an interminable eye illness where optic nerve is continuously harmed and gradually begins to cause sight misfortune [1]- [9].
In its beginning times, there is no torment and patients frequently have no indications. After some time glaucoma begins to influence your side/fringe vision and gradually works its way to the center whenever left undetected. As indicated by World Health Organization (WHO), Glaucoma is the subsequent driving reason for vision misfortune; that adds to roughly 5.2 million instances of visual deficiency (15% of absolute visual impairment cases announced) and can possibly influence ~80 million individuals in the following decade. Until this point in time, there is no remedy for glaucoma. Luckily, it is typically a moderate (corresponding e-mail : dr.manjunath.phd@ieee.org) advancing condition, and in the event that it is distinguished early, it very well may be dealt with effectively. Early discovery is the key for forestalling sight misfortune. It is described by the dynamic degeneration of optic nerve strands and prompts auxiliary changes of the optic nerve head, which is known as optic plate, the nerve fiber layer and a synchronous practical disappointment of the visual field. Movement of the infection prompts loss of vision, which happens step by step over an extensive stretch of time. Fig-1 shows the dynamic visual misfortune brought about by glaucoma [1]- [9].

II. TYPES OF GLAUCOMA
There are a few sorts of glaucoma. The two primary sorts are open-edge and edge conclusion. These are set apart by an expansion of intraocular pressure (IOP), or weight inside the eye.   Neo-vascular glaucoma: Neo-vascular glaucoma is brought about by the anomalous development of fresh recruits vessels on the iris and over the eye's waste channels. Neovascular glaucoma is constantly connected with diabetes. It never happens all alone. The fresh recruits vessels hinder the eye's liquid from leaving through the trabecular meshwork causing an expansion in eye pressure [1]- [9].
Shed Glaucoma: happens when a flaky, dandruff-like material strips off the external layer of the focal point inside the eye. The material gathers in the edge between the cornea and iris and can stop up the waste arrangement of the eye, causing eye strain to rise [1]- [9].
Pigmentery Glaucoma: happens when the color granules that are in the rear of the iris break into the unmistakable liquid created inside the eye. These modest shade granules stream toward the seepage waterways in the eye and gradually stop up them, causing eye strain to rise [1]- [9].

III. DETECTION OF GLAUCOMA
Glaucoma is one of the normal reasons for visual impairment. It is an incessant eye illness that prompts vision misfortune, in which the optic nerve is logically harmed. As the side effects possibly happen when the malady is very cutting-edge, glaucoma is known as the quiet hoodlum of sight. In spite of the fact that glaucoma can't be relieved, its movement can be eased back somewhere near treatment. Early discovery of glaucoma dependent on viable pictures is profoundly required [1]- [9].
Advanced Fundus Image is one of the primary and well known modalities to analyze glaucoma. Since it is conceivable to secure DFIs in a non-obtrusive way which is appropriate for enormous scope screening, DFI has developed as a favored methodology for huge scope glaucoma screening. In a glaucoma screening program, a mechanized framework chooses whether or no indications of suspicious for glaucoma are available in a picture. Just those pictures esteemed suspect by the framework will be passed to ophthalmologists for additional assessment [1]- [9].
Glaucoma finding in the clinical condition includes intraocular pressure estimation, visual-field testing or optic plate assessment on fundus pictures. Despite the fact that intraocular pressure means that glaucoma, its estimation isn't a successful method for glaucoma determined as certain patients to have glaucoma may have typical eye pressure? Visual-field testing, then again, requires extraordinary hardware that a few centers might not have. The last technique, optic plate assessment, is more advantageous than the other two and is all the more generally utilized by masters for early glaucoma recognition [1]- [9].
The nonexclusive computerized glaucoma location process is outlined in Fig. 3. In glaucoma location as a matter of first importance picture of retina is taken utilizing computerized picture catching gadgets. At that point preprocessing is required for evening out of anomalies with pictures. Highlight extraction includes streamlining the measure of assets required to depict a huge informational collection precisely. A component is a critical information that can be utilized for characterization. Characterization alludes to the examination of the properties of a picture. Contingent on the investigation, the dataset is additionally alluded into various classes for example typical or glaucoma affected [1]- [9].

IV. PROBLEM STATEMENT
Glaucoma is a main eye sickness, causing vision misfortune by step by step influencing fringe vision whenever left untreated. Current conclusion of glaucoma is performed by ophthalmologists, human specialists who regularly need to break down various sorts of clinical pictures created by various kinds of clinical gear: fundus, Retinal Nerve Fiber Layer (RNFL), Optical Coherence Tomography (OCT) circle, OCT macula, perimetry, as well as perimetry deviation. Catching and breaking down these clinical pictures is work concentrated and tedious. Consequently we present a novel methodology for glaucoma analysis depending on fundus pictures utilizing profound convolutional neural systems, for example, (1) VGG Networks (VGGNets), • To build up a profound learning (DL) design with convolutional neural system for robotized glaucoma finding utilizing diverse sort of change procedures. • To analyze them for their best execution for glaucoma location by discovering the exhibition lists. CNN is a cutting edge technique, in view of its capacity to separate highlights in pictures without complex prehandling, combined with move learning and tweaking parameters. This examination utilizes VggNet, Alexnet, InceptionNet, GoogleNet and Resnet, which are move adapting frequently utilized in profound learning. We use move figuring out how to get the element vector for arranging diabetic retinopathy utilizing SVM and think about the outcomes, which move learning is the best for characterizing diabetic retinopathy. The grouping layer is evacuated, and the last completely associated layer is applied to get the highlights for the arrangement procedure utilizing the help vector machine (SVM) as appeared in Fig.

Fig. 4. Proposed method design
A CNN engineering by and large comprises of convolutional layers, pooling layers or subsampling layers, completely associated layers, and the order layer as appeared in Fig. 4 [1]- [9].

VIII. CONVOLUTIONAL NEURAL NETWORKS
Given their high viability, CNNs are as of now the most generally utilized method in picture classification . Their quality stems from the utilization of convolutional filters that are made out of neurons, otherwise called open fields. Propelled by natural procedures, the neurons convolve over neighborhood districts in the information layer and react to specific examples, as the visual cortex does to boosts in a nearby space. This is a significant trademark on the grounds that, in contrast to the full availability of common neural systems, this nearby network empowers CNNs to deal with high-dimensional data sources, for example, regular pictures by generously decreasing the quantity of parameters to process. Furthermore, CNNs can learn and arrange progressive highlights legitimately from crude information pictures (start to finish learning), therefore not requiring extraction of hand-made highlights.
As CNNs have been exhibited to be exceptionally compelling across different general picture classification assignments, including the ImageNet Large-Scale Visual Recognition Challenge (ILSVRC), which accompanies more than 1.2 millions pictures dispersed more than 1000 classes [33], clinical picture investigation utilizing CNNs followed rapidly , focusing on use cases, for example, diabetic retinopathy finding and chest pathology examination. To accomplish our objective of compelling glaucoma conclusion, we investigated five delegate CNNs: (1) VGG Networks (VGGNets), Inception Networks (InceptionNets) , and (3) Residual Networks (ResNets), (4) GoogleNet and AlexNet. These systems are completely founded on more profound structures, all accomplishing higher correctnesses while keeping up lower blunder rates. An examination of the exactnesses accomplished by the various kinds of systems are is established [1]- [9].

A. VGG Nets:
Created by Oxford University, VGGnetwas an exchange discovering that progressively proper ConvNet structures, which both acquire the best in class exactness on ILSVRC arrangement and appropriate to another picture acknowledgment datasets [13], and turned into the second champ of the 2014 ILSVRC with a mistake rate of6.8%. VGGnet utilizes an information picture of 224 x 224 with three channels (Red Green Blue). Appeared in Fig. 5, VGGnet has five squares of convolutional forms, with the first blockhas two convolution layers with relu initiation, trailed by a pooling layer, the subsequent square has two convolution layers with relu actuation followed by a pooling layer, the third square has three convolution layers followed a pooling layer, the fourth square has three convolution layers followed by a pooling layer, and the last square has three convolution layers followed a pooling layer, at that point followed by completely associated layers fc6, fc7, and fc8 followed by the softmax layer has 1000neurons class scores. A convolutional layer of Vggnet engineering utilizes an exceptionally little piece size, yet VggNet requires a long registering time [1]- [9].

B. Alexnet
Alexnet was created by Alex Krizhevsky Alexnet is a sort of CNN and is more profound than Lenet. Contrasted with the past CNN technique, Alexnet is progressively ideal in the element extraction process. Alexnet utilizes 227 x 227 information pictures with three channels (Red Green Blue), has five convolutional layers and three max-pooling layers and three completely associated layers. ILSVRC 2012 rivalry Alexnet is a victor with a blunder pace of 16.4%from aninitial mistake of26.2%. Alexnet utilizes the info picture with a size of227 x 227and3channels,the first convolution is trailed by the enactment and bunch standardization with max-pooling layer, the second convolutional layer followed by the actuation and cluster standardization with maxpooling layer, the third convolutional layer followed by the initiation, the fourth convolutional layer followed by the enactment, the fifth convolutional layer followed by initiation of relu and max-pooling layer, completely associated layers comprise offc6, fc7, and fc8, trailed by softmax and the order layer which has 1000 neurons as thenumber of classes [1]- [9].

C. InceptionNet
Christian Szegedy created systems brought in subtitle net. Commencement has an info picture with a size of 299 x 299 and three channels (Red Green Blue). With three 154 convolutional layers and pooling layers in square 1 and convolution layers in square two followed by three initiation hinders for lessening parameter, at that point the pooling layer and arrangement [1]- [9].

D. Resnet
Kaiming created Resnet (Residual Neural Network) move learning [15] for the ILSVRC rivalry 2015. Resnet utilized leftover squares effectively prepared 152 layers with a blunder pace of 4.49% for a solitary model on the ImageNet approval set, and 3.57% top-5 mistake on the test set for the group of six diverse profundity models. Resnet has better execution despite the fact that it has lower multifaceted nature than VGGnet. Resnet organize utilizes an information picture of size 224 x 224 with three channels (Red Green Blue), convolution and pooling. This investigation utilizes the completely associated layer which has a vector highlight to be utilized in the order arrange utilizing the help vector machine [1]- [9].

E. GoogleNet
Googlenet is an exchange discovering that executes an initiation structure that can keep up computational execution. In the ILSVRC rivalry, Googlenet utilizes 22 CNN layers which is progressively ideal contrasted with VGGnet, GoogleNet produces 5.5% blunders with top-5 characterizations. The design of GoogleNet is mind boggling by including initiation modules in the engineering. GoogleNet utilizes input picture with the size of 224 x 224 and 3 channels (red, green, blue), with origin layers, each square ofan commencement layer comprises of 6 convolution layers and a pooling layer that can diminish the profound of the picture include map, at that point completely associated misfortune f-classifier layer, GoogleNet has the design with tedious preparing due to the more profound layers, yet GoogleNet is more exact than Alexnet and VGGNet. This examination utilizes the loss3classifier layer as a component vector from GoogleNet to be utilized as an element vector in the characterization procedure utilizing SVM [1]- [9].

IX. DATASET: FUNDUS IMAGES
The fundus informational index utilized for preparing contained 540 typical eye pictures and 1363 glaucomatouseye pictures, in this manner bringing about the utilization of a sum of 1903 eye pictures. This informational collection was haphazardly part into two sets, to be specific a preparation and an approval set, utilizing a split proportion of roughly 4:1(that is, the preparation dataset contained 1503 pictures and the approval dataset 400 pictures). The preparation set was utilized to fine-tune our models, though the approval set was utilized to evaluate the speculation intensity of the models made during preparing. Aside from these two datasets, for testing purposes, 220 concealed fundus pictures were utilized: 55 ordinary eye pictures and 165 glaucomatous eye pictures. The pictures utilized accompanied diverse spatial goals, running from 1172-2500 pixels on a level plane and from 1500-3200 pixels vertically [1]- [9].

X. IMPLEMENTATION
Our general trial arrangement has streaming stages, which are clarified as follows [1]- [9].
Information pre-preparing: In this stage, all the information pictures were re-sized and standardized before preparing, considering the pre-prepared model settings. Information expansion: Over fitting happens every now and again when joining a little measured dataset with a profound neural system, similar to the case for the given arrangement of fundus pictures and the models utilized, with the last accompanying an enormous number of parameters that should have been figured. So as to lighten over fitting, we consolidated information enlargement, artificially enhancing the volume of information during preparing through picture handling procedures (e.g., level flipping, differentiate change, and splendor modification), and where these strategies are frequently utilized with regards to PC vision assignments. In light of an examination of the precision of various information increase systems in, we applied a few irregular information growth techniques, likewise considering the various models utilized (that is, for each model, we utilized various sorts of information enlargement) [1]- [9].
Adjusting: When preparing without any preparation, a model regularly introduces all parameters utilizing arbitrary Gaussian conveyances. This methodology towards introduction is generally less exact, with the current model requiring more opportunity to unite. To stay away from introduction issues, we utilized models pre-prepared on ImageNet, moving their loads to our models, thusly additionally taking into account better speculation. Beginning from these pre-prepared loads, we fine-tuned all layers utilizing the hyper parameter esteems [1]- [9].

XI. EVALUATION
Other than approval during the preparation stage, we actualized two additional ways to deal with assess our model speculation. In the first place, as clarified in Section III, we made another dataset for testing purposes as it were. Second, we utilized RIM-ONE r3, an openly accessible informational index of fundus pictures, to assess our model generalizability. Assessment results got for both datasets are accounted for in the accompanying segment utilizing a few measurements: exactness, review (affectability), accuracy, F1 score, specificity, and Receiver Operating Characteristic-Area Under the Curve (ROC-AUC). These measurements were determined as follows [1]- [9] XII. LEARNING MODEL EVALUATION CRITERIA : The precision, affectability, particularity, probability proportion, and ROC/AUC have been generally utilized as criteria for assessing a conclusion model. The accompanying terms are central to understanding the utility of them: [1]- [9] 1. True positive (TP): the patient has an ailment and the forecast is certain. The exactness of a determination model alludes to the capacity of the model to effectively distinguish those patients with the illness and without the infection: [1]- [9] The affectability of a determination model alludes to the capacity of the model to effectively recognize those patients with the malady: [1]- [9] The particularity of a conclusion model alludes to the capacity of the test to effectively distinguish those patients without the ailment: [1]- [9] The probability proportion is characterized as the proportion of expected test brings about subjects with a specific infection to the subjects without the ailment. The Likelihood proportion for positive test outcomes (LR+) reveals to us the amount more probable the positive test outcome is to happen in subjects with the ailment contrasted with those without the malady: [1]- [9] The Likelihood proportion for negative test outcomes (LR-) speaks to the proportion of the likelihood that a negative outcome will happen in subjects with the illness to the likelihood that a similar outcome will happen in subjects without the malady: [1]- [9] The collector working trademark (ROC) plot communicates connection among affectability and 1 -Specificity. The closer the ROC bend is situated to upperleft hand corner, the better the model. The region under the curve(AUC) can have any an incentive somewhere in the range of 0 and 1 and it is a decent pointer of the integrity of the model. [1]- [9] XIII. SOFTWARE AND HARDWARE TOOL REQUIRED

A. Software tools used:
Ongoing works DL-based strategies for visual ailments are constantly executed utilizing TensorFlow library, either with or without Keras API, Caffe system or Matlab. The three arrangements have programming instruments to consequently play out the preparation on GPU models so as to profit by equal handling execution and progressive memory with the point of diminishing the preparation execution time. The Matlab instrument permits actualizing all means required by the DL-strategies both for prepreparing and preparing. Nonetheless, even TensorFlow and Caffe protect information enlargement and squaring handling, they are restricted to perform pre-preparing to improve input picture, for example, applying channels, morphologic administrators, and so forth. Right now, works continue to join the OpenCV library that speaks to an Open Computer Vision broadly written in C++. [1]- [9] The software environment that can be used for implementation are : • Keras API for TensorFlow library The hardware kits that are going to be used for the project work is DSP card with CCS-Code Composer Studio & the National Instruments Kits / FPGA Spartan kits for the experimentation purposes / Xilinx Spartan Kits or any other hardware interfacing platform. [1]- [9] XIV. POSSIBLE OUTCOME Glaucoma detecting framework was depended on the deep-learning C N Ns and was able to detect the characterizing features that would better show the hidden pattern which were related to the glaucomatic disease. The framework in the DL concept was implemented with the XV. CONCLUSION As interactive media information is developing exponentially, further developed methods are expected to deal with such immense measure of unstructured information. Right now, study the upsides of using profound learning systems recognize glaucoma utilizing fundus pictures. The DL structure for glaucoma is executed utilizing profound convolutional neural systems, for example, AlexNet, GoogLeNet, VGG, ResNet and DenseNet. DL, a thriving innovation of ML, can find complicated structures in informational indexes without the need to determine manages expressly. A DL organize is a CNN with numerous layers between the info and yield layers. It has drastically improved the best in class in picture acknowledgment.